/external/XNNPACK/src/f32-dwconv/gen/ |
D | up8x4-wasmsimd.c | 66 const v128_t vk0x0123 = wasm_v128_load(w + 8); in xnn_f32_dwconv_ukernel_up8x4__wasmsimd() local 114 const v128_t vk0x0123 = wasm_v128_load(w + 8); in xnn_f32_dwconv_ukernel_up8x4__wasmsimd() local 147 const v128_t vk0x0123 = wasm_v128_load(w + 8); in xnn_f32_dwconv_ukernel_up8x4__wasmsimd() local
|
D | up8x4-minmax-wasmsimd-arm.c | 68 const v128_t vk0x0123 = wasm_v128_load(w + 8); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm() local 119 const v128_t vk0x0123 = wasm_v128_load(w + 8); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm() local 153 const v128_t vk0x0123 = wasm_v128_load(w + 8); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm() local
|
D | up8x4-minmax-wasmsimd-x86.c | 68 const v128_t vk0x0123 = wasm_v128_load(w + 8); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86() local 119 const v128_t vk0x0123 = wasm_v128_load(w + 8); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86() local 153 const v128_t vk0x0123 = wasm_v128_load(w + 8); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86() local
|
D | up8x4-minmax-sse.c | 68 const __m128 vk0x0123 = _mm_load_ps(w + 8); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse() local 118 const __m128 vk0x0123 = _mm_load_ps(w + 8); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse() local 152 const __m128 vk0x0123 = _mm_load_ps(w + 8); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse() local
|
D | up8x4-minmax-sse-acc2.c | 68 const __m128 vk0x0123 = _mm_load_ps(w + 8); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse_acc2() local 121 const __m128 vk0x0123 = _mm_load_ps(w + 8); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse_acc2() local 157 const __m128 vk0x0123 = _mm_load_ps(w + 8); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse_acc2() local
|
D | up8x4-minmax-neon.c | 67 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x4__neon() local 107 const float32x4_t vk0x0123 = vld1q_f32(w + 4); in xnn_f32_dwconv_minmax_ukernel_up8x4__neon() local 133 const float32x4_t vk0x0123 = vld1q_f32(w + 8); in xnn_f32_dwconv_minmax_ukernel_up8x4__neon() local
|
D | up8x4-minmax-wasmsimd-arm-acc2.c | 68 const v128_t vk0x0123 = wasm_v128_load(w + 8); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm_acc2() local 122 const v128_t vk0x0123 = wasm_v128_load(w + 8); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm_acc2() local 158 const v128_t vk0x0123 = wasm_v128_load(w + 8); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm_acc2() local
|
D | up8x4-minmax-neonfma.c | 67 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma() local 107 const float32x4_t vk0x0123 = vld1q_f32(w + 4); in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma() local 133 const float32x4_t vk0x0123 = vld1q_f32(w + 8); in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma() local
|
D | up8x4-minmax-wasmsimd-x86-acc2.c | 68 const v128_t vk0x0123 = wasm_v128_load(w + 8); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86_acc2() local 122 const v128_t vk0x0123 = wasm_v128_load(w + 8); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86_acc2() local 158 const v128_t vk0x0123 = wasm_v128_load(w + 8); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86_acc2() local
|
D | up8x4-minmax-neon-acc2.c | 67 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2() local 110 const float32x4_t vk0x0123 = vld1q_f32(w + 4); in xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2() local 138 const float32x4_t vk0x0123 = vld1q_f32(w + 8); in xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2() local
|
D | up8x4-minmax-neonfma-acc2.c | 67 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2() local 110 const float32x4_t vk0x0123 = vld1q_f32(w + 4); in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2() local 138 const float32x4_t vk0x0123 = vld1q_f32(w + 8); in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2() local
|
D | up4x4-wasmsimd.c | 64 const v128_t vk0x0123 = wasm_v128_load(w + 4); in xnn_f32_dwconv_ukernel_up4x4__wasmsimd() local 97 const v128_t vk0x0123 = wasm_v128_load(w + 4); in xnn_f32_dwconv_ukernel_up4x4__wasmsimd() local
|
D | up4x4-minmax-sse.c | 66 const __m128 vk0x0123 = _mm_load_ps(w + 4); in xnn_f32_dwconv_minmax_ukernel_up4x4__sse() local 100 const __m128 vk0x0123 = _mm_load_ps(w + 4); in xnn_f32_dwconv_minmax_ukernel_up4x4__sse() local
|
D | up4x4-minmax-wasmsimd-x86.c | 66 const v128_t vk0x0123 = wasm_v128_load(w + 4); in xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_x86() local 101 const v128_t vk0x0123 = wasm_v128_load(w + 4); in xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_x86() local
|
D | up4x4-minmax-wasmsimd-arm.c | 66 const v128_t vk0x0123 = wasm_v128_load(w + 4); in xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_arm() local 101 const v128_t vk0x0123 = wasm_v128_load(w + 4); in xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_arm() local
|
D | up4x4-minmax-wasmsimd-arm-acc2.c | 66 const v128_t vk0x0123 = wasm_v128_load(w + 4); in xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_arm_acc2() local 103 const v128_t vk0x0123 = wasm_v128_load(w + 4); in xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_arm_acc2() local
|
D | up4x4-minmax-sse-acc2.c | 66 const __m128 vk0x0123 = _mm_load_ps(w + 4); in xnn_f32_dwconv_minmax_ukernel_up4x4__sse_acc2() local 102 const __m128 vk0x0123 = _mm_load_ps(w + 4); in xnn_f32_dwconv_minmax_ukernel_up4x4__sse_acc2() local
|
D | up4x4-minmax-neon.c | 65 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x4__neon() local 91 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x4__neon() local
|
D | up4x4-minmax-wasmsimd-x86-acc2.c | 66 const v128_t vk0x0123 = wasm_v128_load(w + 4); in xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_x86_acc2() local 103 const v128_t vk0x0123 = wasm_v128_load(w + 4); in xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_x86_acc2() local
|
D | up4x4-minmax-neonfma.c | 65 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma() local 91 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma() local
|
D | up4x4-minmax-neonfma-acc2.c | 65 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma_acc2() local 93 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma_acc2() local
|
D | up4x4-minmax-neon-acc2.c | 65 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x4__neon_acc2() local 93 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x4__neon_acc2() local
|
D | up8x9-wasmsimd.c | 91 const v128_t vk0x0123 = wasm_v128_load(w + 8); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local 184 const v128_t vk0x0123 = wasm_v128_load(w + 8); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local 247 const v128_t vk0x0123 = wasm_v128_load(w + 8); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local
|
D | up4x9-wasmsimd.c | 89 const v128_t vk0x0123 = wasm_v128_load(w + 4); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd() local 152 const v128_t vk0x0123 = wasm_v128_load(w + 4); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd() local
|
D | up8x9-minmax-sse.c | 93 const __m128 vk0x0123 = _mm_load_ps(w + 8); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local 188 const __m128 vk0x0123 = _mm_load_ps(w + 8); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local 252 const __m128 vk0x0123 = _mm_load_ps(w + 8); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local
|